From 386614fa68bfd79313ea965da65bad18b2d2fbce Mon Sep 17 00:00:00 2001
From: Fraser <tomatoesneverlie@gmail.com>
Date: Mon, 3 Apr 2023 20:20:36 -0400
Subject: [PATCH] Finish restructure prompt

I've A/B tested *a lot*, and find doing it this way (sources initially, then truncated conversation, then final question) ends up allowing the user to ask clarifying questions better, with minimal if any drop in citation knowledge and relevance
---
 api/chat.py             | 15 ++++++++++++++-
 web/src/pages/index.tsx |  8 +++-----
 2 files changed, 17 insertions(+), 6 deletions(-)

diff --git a/api/chat.py b/api/chat.py
index 0284bb8..f948601 100644
--- a/api/chat.py
+++ b/api/chat.py
@@ -4,8 +4,8 @@ from get_blocks import get_top_k_blocks, Block
 
 from typing import List, Dict
 import openai
-import os
 import tiktoken
+import re
 
 # OpenAI models
 EMBEDDING_MODEL = "text-embedding-ada-002"
@@ -70,9 +70,18 @@ def construct_prompt(query: str, history: List[Dict[str, str]], context: List[Bl
             source_prompt += block_str
             token_count += block_tc
 
+    source_prompt = source_prompt.strip();
+    if len(history) > 0:
+        source_prompt += "\n\n"\
+            "Before the question (\"Q: \"), there will be a history of previous questions and answers. " \
+            "These sources only apply to the last question. Any sources used in previous answers " \
+            "are invalid."
+
     prompt.append({"role": "system", "content": source_prompt.strip()})
 
 
+
+
     # Write a version of the last 10 messages into history, cutting things off when we hit the token limit.
     token_count = 0
     history_trnc = []
@@ -82,6 +91,10 @@ def construct_prompt(query: str, history: List[Dict[str, str]], context: List[Bl
             token_count += len(ENCODER.encode("Q: " + message["content"]))
         else:
             content = cap(message["content"], int(NUM_TOKENS * HISTORY_FRACTION) - token_count)
+
+            # censor all source letters into [x]
+            content = re.sub(r"\[[0-9]+\]", "[x]", content)
+
             history_trnc.append({"role": "assistant", "content": content})
             token_count += len(ENCODER.encode(content))
 
diff --git a/web/src/pages/index.tsx b/web/src/pages/index.tsx
index ec90794..3c6712c 100644
--- a/web/src/pages/index.tsx
+++ b/web/src/pages/index.tsx
@@ -23,7 +23,6 @@ type UserEntry = {
 type AssistantEntry = {
     role: "assistant";
     content: string;
-    display_content: string;
     citations: Map<number, Citation>;
 }
 
@@ -89,7 +88,7 @@ const ShowEntry: React.FC<{entry: Entry}> = ({entry}) => {
     return (
         <div className="mt-3 mb-8">
             {   // split into paragraphs
-                entry.display_content.split("\n").map(paragraph => ( <p> {
+                entry.content.split("\n").map(paragraph => ( <p> {
                     paragraph.split(in_text_citation_regex).map((text, i) => {
                         if (i % 2 === 0) {
                             return text.trim();
@@ -143,7 +142,7 @@ const Home: NextPage = () => {
                            .map((entry) => {
                                return {
                                    "role" : entry.role,
-                                   "content" : entry.content
+                                   "content" : entry.content.trim(),
                                }
                            })
                })
@@ -234,8 +233,7 @@ const Home: NextPage = () => {
         });
 
         setEntries([...new_entries, {role: "assistant", 
-                                     content: await data.response, 
-                                     display_content: response,
+                                     content: response,
                                      citations: citations}]);
 
         setLoading(false);